Admissions, not craving scores: what a Swedish register found for four alcohol medicines
- Nationwide observational register study from Karolinska Institutet, with co-authors in Mannheim, Helsinki and Turku, using the Swedish National Patient Register, the Prescribed Drug Register, the Causes of Death Register and the LISA database. All residents aged 18–64 with a first-time alcohol-related diagnosis (ICD-10 F10.0–F10.9) in inpatient or specialised outpatient care between 2009 and 2019 were included; anyone with an alcohol diagnosis between 1997 and 2008, or with an earlier prescription for an alcohol medicine, was excluded. N = 93,727, of whom 57,730 (61.6%) were men, mean age 37.0 years (standard deviation 14.6), median follow-up 5.0 years. Entry diagnosis was acute intoxication (F10.0) for 44,819 (47.8%) and dependence (F10.2) for 23,340 (24.9%). Over the follow-up, 35,292 people (37.7%) had at least one alcohol-related admission.
- The primary outcome declared in the Methods is an inpatient admission with a main diagnosis of F10.0–F10.9, treated as a recurrent event in a between-individual Cox model with time since the last admission as the time scale and robust standard errors. Exposure to any of the four medicines was associated with a lower risk of admission than periods without exposure: hazard ratio 0.85 (95% CI 0.81 to 0.88), p < 0.0001. As monotherapy, naltrexone 0.80 (95% CI 0.73 to 0.87), p < 0.0001, and disulfiram 0.83 (95% CI 0.79 to 0.88), p < 0.0001; acamprosate 0.96 (95% CI 0.89 to 1.03), p = 0.2224, and nalmefene 0.79 (95% CI 0.45 to 1.39), p = 0.4095, did not reach significance. Combinations: disulfiram with acamprosate 0.57 (95% CI 0.44 to 0.74), p < 0.0001; naltrexone with disulfiram 0.68 (95% CI 0.49 to 0.96), p = 0.0267; naltrexone with acamprosate 0.82 (95% CI 0.67 to 1.00), p = 0.0509.
- Nobody rated the outcome. An admission is a record in a national register, so there is no self-report, no clinician impression and no blinding to arrange. The surrogate in this study sits on the exposure side instead: a person counts as treated from the date a prescription was dispensed until the purchased defined daily doses run out, which measures pharmacy pick-up rather than swallowing. The authors state the assumption and argue it biases toward the null.
- Death was available and was not analysed. 4,565 people (4.9%) died during follow-up; the Causes of Death Register supplied dates of death, and death served to end observation rather than to count as an outcome. There is no mortality, no overdose and no return-to-work analysis in this paper.
Alcohol pharmacotherapy is usually reported in units that never leave the questionnaire: drinking days, craving scores, grams per day. This study went to a national register and counted the thing that costs a bed, asking whether periods covered by naltrexone, acamprosate, disulfiram or nalmefene carried a different risk of an alcohol-related admission than periods without.
The primary outcome is declared in the Methods and is exactly what the abstract reports, so no outcome switching is visible, though the study is observational and cites no trial registration or published protocol. The model adjusts for sex, education, age, number of previous exposure periods, previous alcohol-related admissions, cohabitation, comorbid substance use disorders (F11–F19), somatic comorbidity and benzodiazepine exposure, but not for motivation, refusal of medication or psychosocial treatment received alongside, and the authors say so.
Two places where the abstract runs ahead of the results
The abstract says that stratified subgroup analyses "confirmed the robustness of the results". For age the results section does not read that way. Among people aged 18–24 the association ran the other way and was significant, 1.19 (95% CI 1.03 to 1.39), p = 0.0201; among those aged 25–34 it did not reach significance, 1.03 (95% CI 0.94 to 1.12), p = 0.5884; only from 35 years upwards does the range 0.74 to 0.85, lying below one, appear. Neither the youngest group nor the reversal reaches the abstract, and the discussion mentions younger patients only in connection with prescription rates, not with any change in the direction of the association. I followed the results section.
Most of the stratification did hold, and reporting only the failures would distort the paper the other way. The association stayed significant in men, 0.82 (95% CI 0.78 to 0.86), and women, 0.90 (0.84 to 0.97); with a partner, 0.79, and alone, 0.87; across every stratum of somatic comorbidity, 0.78 to 0.85; and in the primary, secondary and university education groups, 0.82 to 0.90. The sharper objection is that the one socioeconomic stratum with no association is the very category the abstract invokes, unknown or no completed primary education, 0.82 (95% CI 0.63 to 1.06), p = 0.1277, so the robustness claim fails on its own narrow ground. The sensitivity analysis in the 23,340 people who entered with a dependence diagnosis did reproduce the significant associations for any medication, naltrexone, disulfiram and disulfiram with acamprosate. The full set of subgroup estimates sits in supplementary Figure S1, not in the text.
The second place is the within-individual analysis, where each person is their own control and time-invariant confounders drop out. It included 8,232 people, 8.8% of the cohort, since only those exposed at least once and admitted at least once contribute. There the naltrexone with disulfiram combination did not reach significance, while any medication, naltrexone, disulfiram and disulfiram with acamprosate did. The text calls this corroboration and reports the failure in the next sentence, the abstract carries that combination without qualification, and the discussion calls the combinations synergistic. Its hazard ratios are read off Figure 3 and appear numerically nowhere in the main text or figures; the supplementary Data S1 was not opened.
Acamprosate and nalmefene were not measured on their own terms
Nalmefene was dispensed to 547 people, 0.6% of the cohort, and its interval runs from 0.45 to 1.39: the authors put that down to the low number of events, an absence of precision rather than of effect, and note that nalmefene is usually prescribed for harm reduction rather than abstinence, so admission is a mismatched endpoint for it. For acamprosate their explanation is that it may do less against the heavy drinking episodes that produce intoxication admissions. Concluding that two of the four medicines do not work would be the symmetrical error.
The authors put the hazard ratios first
Their headline is the hazard ratios. The Aims of the Study section names the association between medication exposure and hospitalisation risk as the primary aim and prescription rates as a secondary one; in the Significant outcomes box underuse is the third of three bullets; the abstract's Conclusion opens with "Results indicate a significant association between disulfiram and naltrexone monotherapy ... with a lower risk", and the Discussion opens the same way. Their own Conclusion goes further: combinations "show even stronger associations", and the findings "highlight the protective effect of AUD medication" – the causal word is theirs, not ours – before it turns to prescribing. Underuse is their closing call to action, not their finding.
The sharpest arithmetic in the paper is nonetheless there. Only 23,649 people (25.2%) were ever dispensed one of these medicines, and 15,941 (17.0%) within the first year after diagnosis, and the gradient inside that share is given in Table 2 rather than in the text: 5% of 18–24 year olds against 27% of 45–54 year olds, 10.2% of those with no completed primary education against 22.3% of university graduates, 4.1% of those who entered on an intoxication code against 34.4% of those diagnosed with dependence. Disulfiram was the most prescribed of the four across the whole follow-up, reaching 12,113 people (12.9%) against 10,319 (11.0%) for naltrexone. The paper is not the first of its kind and does not claim to be: it cites a Swedish register cohort of 125,556 patients (Heikkinen and colleagues, Addiction, 2021) reporting the same direction for naltrexone and disulfiram.
What changes in the room
You can now say that in a whole national cohort, periods covered by naltrexone or disulfiram carried a hazard of alcohol-related admission roughly a fifth lower than uncovered periods, on an outcome no patient and no clinician rated. You cannot say that the drugs caused it, that they reduce death, or that a young patient will show the same association. Funding was public and no conflict of interest is declared. What remains missing is the outcome one step beyond the one they chose: the deaths were in the file.
The outcome here is an admission recorded by a national register rather than anything a patient or a clinician rated, and the deaths of 4,565 people in the cohort were used to end observation rather than counted as an outcome.
This is observational register data with no trial registration or published protocol, so association cannot be read as causation; exposure was estimated from dispensed defined daily doses, which measures pharmacy pick-up rather than actual intake. Treatment motivation, family history and concurrent psychosocial treatment were unavailable, and confounding by indication cannot be excluded in either direction. The within-individual analysis, which controls time-invariant confounding, included only 8,232 of 93,727 people and did not confirm the naltrexone with disulfiram combination. No correction for multiple comparisons is described for the subgroup analyses, so the reversed association among 18–24 year olds is a signal to check rather than a finding to act on. Mortality and overdose were not analysed.